FAR AI researchers show linear probes can catch AI sabotage

Researchers at FAR AI, led by founder Adam Gleave, published a paper arguing that lightweight linear probes reading a model’s internal activations can scale up to catch sabotage and deception that text-based monitoring misses. The team built what they described as the largest deception dataset assembled to date, FIBS, and introduced a probe architecture able to aggregate signal across many layers and tokens of a model’s internal state, rather than reading only its final output.

On SHADE-Arena, a benchmark for detecting models secretly sabotaging a task, the probes reached 98.8% AUC, outperforming a baseline that used Claude Opus 5.5 itself as a text-based monitor, and the researchers reported the probes’ accuracy improved further as the underlying model was scaled up. They also tested the method on cases they called “introspective deception,” where whether a model is lying cannot be determined from its output text alone and ground truth is only recoverable through careful elicitation or knowledge of its training data; in one such test, probes distinguished transcripts containing a model’s genuine hidden goal from decoy goals with up to 99.7% AUC. The probes also detected deception in prominent open-weight models that misrepresented their views on politically sensitive topics or lied under pressure about their own beliefs.

The result adds to a small but growing body of work arguing that interpretability-based, “white-box” monitoring — reading a model’s internal states directly — can outperform asking another model to read its outputs, particularly for behaviour a model does not put into words at all. The group released its FIBS training dataset publicly, saying it hoped to encourage wider deployment of probe-based monitoring and further work expanding the dataset with new examples of deception and sabotage.